COMBINING MODEL-BASED AND IN SITU PERFORMANCE PREDICTION TO EVALUATE DETECTION CLASSIFICATION PERFORMANCE

摘要

In conclusion, Manning has previously demonstrated that classification performance can be accurately predicted by SNR_(eff) in Rayleigh-distributed backgrounds. The work presented herein extends this result by accounting for a wide variety of background statistics as captured by the shape parameter of the k-distribution. It also presents a real time tool (or offline simulation) capable of accurate performance prediction using either models and a prior specifications or actual imagery gathered in situ. The extremely low computational complexity and cost of this technique stems from the fact that the mapping between the key parameters and performance is learned offline and stored for use in real time. Thus, these key parameters may be estimated in situ (or derived from a model) and immediately combined to produce an accurate performance estimate. The future research of this work concern more robust estimation techniques for shadow contrast and roughness. Although not discussed, the primary method for ascribing roughness to an area is via (US) Naval Warfare Publication NWP 3-15.41. While this method is used because it is US Navy Doctrine, higher fidelity techniques are under investigation.